Glossary

context window

The slice of text a model can see at one moment: its working memory, measured in tokens.

What it is

A large language model reads its whole window from scratch on every turn, predicts the next stretch of text, and stops. Everything inside the window, your prompt, the files you pasted, the model's own earlier replies, can shape the answer. Everything outside it might as well not exist. The window is measured in tokens, chunks of text roughly the size of a short word. Think desk, not hard drive: only what's on the desk gets used. The thing you need ends up off the desk one of two ways. Either you never put it there, or you put it there early and the conversation pushed it off the top. Either way, the model fills the gap by guessing.

Why CEOs care

This is the real reason a long task degrades and the reason a "smarter model" often isn't the fix. An agent running a multi-step job crowds its own window with tool output and half-finished steps until early decisions scroll off, which is why something sharp in a five-minute demo falls apart over a multi-hour job. Bigger windows aren't a silver bullet either: models degrade as input grows, well before the maximum, an effect known as "context rot." The lever is managing what the model sees, not renting a fancier one.

Where you'll see it

A chat that "forgets" what you said earlier in the same conversation once it's gotten long.
An agent that loses the plot on a long job and finishes confidently in the wrong direction.
A "context limit reached" message, or replies that get vaguer as a thread runs long.

Full read

For the CEO-length version, see What is a context window?.

Workflows that use this

Go deeper

Put this term to work.

The Desk Theory guides turn definitions like this into running workflows. Two operator manuals, $99 each, or the bundle for $149.

Browse the books →

← All terms